PLOS ONE
RESEARCH ARTICLE
a1111111111
a1111111111
a1111111111
a1111111111
a1111111111
OPEN ACCESS
Citation: Varrecchia T, Castiglia SF, Ranavolo A,
Conte C, Tatarelli A, Coppola G, et al. (2021) An
artificial neural network approach to detect
presence and severity of Parkinson’s disease via
gait parameters. PLoS ONE 16(2): e0244396.
https://doi.org/10.1371/journal.pone.0244396
Editor: J. Lucas McKay, Emory University, UNITED
STATES
Received: June 17, 2020
Accepted: December 8, 2020
Published: February 19, 2021
Copyright: © 2021 Varrecchia et al. This is an open
access article distributed under the terms of the
Creative Commons Attribution License, which
permits unrestricted use, distribution, and
reproduction in any medium, provided the original
author and source are credited.
Data Availability Statement: All relevant data are
within the paper.
Funding: The research presented in this article was
carried out as part of the SOPHIA project, which
has received funding from the European Union’s
Horizon 2020 research and innovation programme
under Grant Agreement No. 871237.
Competing interests: NO authors have competing
interests.
PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021
An artificial neural network approach to
detect presence and severity of Parkinson’s
disease via gait parameters
Tiwana Varrecchia ID1*, Stefano Filippo Castiglia2, Alberto Ranavolo1, Carmela Conte3,
Antonella Tatarelli1,4, Gianluca Coppola2, Cherubino Di Lorenzo2, Francesco Draicchio1,
Francesco Pierelli2, Mariano Serrao2
1 Department of Occupational and Environmental Medicine, Epidemiology and Hygiene, INAIL, Monte Porzio
Catone Rome, Rome, Italy, 2 Department of Medico-Surgical Sciences and Biotechnologies, University of
Rome Sapienza, Latina, Italy, 3 IRCSS Fondazione Don Carlo Gnocchi, Milan, Italy, 4 Department of Human
Neurosciences, University of Rome Sapienza, Rome, Italy
* t.varrecchia@inail.it
Abstract
Introduction
Gait deficits are debilitating in people with Parkinson’s disease (PwPD), which inevitably
deteriorate over time. Gait analysis is a valuable method to assess disease-specific gait pat-
terns and their relationship with the clinical features and progression of the disease.
Objectives
Our study aimed to i) develop an automated diagnostic algorithm based on machine-learn-
ing techniques (artificial neural networks [ANNs]) to classify the gait deficits of PwPD
according to disease progression in the Hoehn and Yahr (H-Y) staging system, and ii) iden-
tify a minimum set of gait classifiers.
Methods
We evaluated 76 PwPD (H-Y stage 1–4) and 67 healthy controls (HCs) by computerized
gait analysis. We computed the time-distance parameters and the ranges of angular motion
(RoMs) of the hip, knee, ankle, trunk, and pelvis. Principal component analysis was used to
define a subset of features including all gait variables. An ANN approach was used to iden-
tify gait deficits according to the H-Y stage.
Results
We identified a combination of a small number of features that distinguished PwPDs
from HCs (one combination of two features: knee and trunk rotation RoMs) and identified
the gait patterns between different H-Y stages (two combinations of four features: walking
speed and hip, knee, and ankle RoMs; walking speed and hip, knee, and trunk rotation
RoMs).
1 / 16
PLOS ONE Neural Networks for Parkinson classification

Conclusion
The ANN approach enabled automated diagnosis of gait deficits in several symptomatic
stages of Parkinson’s disease. These results will inspire future studies to test the utility of
gait classifiers for the evaluation of treatments that could modify disease progression.

1. Introduction
Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by a broad
spectrum of motor and non-motor features [1, 2]. Gait deficits represent one of the most debil-
itating aspects in people with Parkinson’ s disease (PwPD); these inevitably decline over the
course of the disease, strongly increase the risk of falls, and greatly reduce patient autonomy
and quality of life [3–5]. The mechanism underlying gait impairment is complex and multifac-
torial and is caused by multi-system lesions involving both the dopaminergic and non-dopa-
minergic mechanisms related to bradykinesia, rigidity, impaired balance and postural control,
visual motor deficiency, and cognition [6, 7]. Given the importance of autonomous and effec-
tive locomotion in humans [8] and considering the socio-economic burden of gait deficits [9],
clinician should strive to optimize pharmacological treatments and rehabilitation interven-
tions to improve gait function.
In recent years, gait analysis has become an essential tool for objective evaluation of gait
changes induced by pharmacological and rehabilitative interventions [10]. Furthermore, it is a
valuable technique to classify gait patterns according to a specific disease or group of diseases
[3, 11–13] or the disease severity [14], allowing more accurate assessment of the quantitative
gait measures to the qualitative clinical features for clinical practice purposes. An innovative
approach to address this issue is the use of quantitative machine-learning techniques such as
artificial neural networks (ANNs), which are mathematical models that represent a distributed
adaptive system built using multiple interconnecting processing elements, just as real neural
networks do [15, 16]. In this model, the processing elements (neurons) are distributed in sev-
eral layers: each neuron receives signals processed and transmitted by neurons in the preced-
ing layer and, in turn, processes and transmits them to the next layer [15, 16].
ANNs have been used in many research fields such as psychology, robotics, biology, com-
puter science, and ergonomics [17, 18] and, more recently, as diagnostic tools in several clini-
cal conditions, including colon or colorectal cancer [19, 20], multiple sclerosis [21, 22],
pancreatic disease [23], gynecological diseases [24], and early diabetes [25]. However, few stud-
ies have attempted to identify and classify gait deficits using machine-learning approaches in
neurological disorders, including Huntington disease [26] and PD [27–29]. Particularly, with
regard to PwPD, most of the published studies investigated two-group gait pattern classifica-
tions, differentiating PwPD from healthy subjects [28, 30, 31], or performed multiclass classifi-
cation according to the disease severity using the Unified Parkinson’s Disease Rating Scale [32,
33]. However, none of these previous studies specifically searched for those gait parameter fea-
tures able to categorize the gait pattern according to disease progression using the Hoehn and
Yahr (H-Y) staging system [34].
The H-Y scale has several strengths and is considered the reference standard for disability
and impairment measurements [35]. It is significantly correlated with dopaminergic loss [35,
36], cerebral spinal fluid and serum alpha-synuclein levels [37], quality of life measurements
[38], and motor performance [39]. Changes in the H-Y stage carry prognostic significance and
influence clinician-based interventions [40]. For instance, the European Guidelines on PD

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 2 / 16

PLOS ONE Neural Networks for Parkinson classification

rehabilitation suggest differentiating the rehabilitation program according to the disease pro-
gression; for instance, promoting balance and gait training for H-Y stages 2 to 4 [41]. There-
fore, identifying the specific gait pattern for each disease stage may allow monitoring of the
changes induced by pharmacological and rehabilitation treatment in each stage, taking into
account only those meaningful parameters able to classify gait deficit in PwPD. When studying
gait features, it is classically adopted a univariate approach, whereby measurement outcomes
are considered independently. This data redundancy may be highly time–consuming and
increase the risk of losing important information [42]. In this view, machine learning
approaches may overcome this limitation, allowing to explore the optimal combination of gait
characteristics reducing the computational demand at the same time [42, 43].
Our study aimed to i) develop a diagnostic algorithm based on machine-learning tech-
niques (i.e., ANNs) able to classify the gait deficit of PwPD according to the disease progres-
sion as evaluated by the H-Y staging system and ii) identify the minimum set of gait time-
distance and kinematic parameters able to distinguish the H–Y stage gait pattern from each
other.

2. Materials and methods

2.1 Patients
This study enrolled 76 PwPD (age: 69.68±8.92 years). Clinical evaluation of the severity of Par-
kinsonism included neurological and functional assessments using the UPDRS III (18±12)
and the H-Y staging system (20 with H-Y = 1 (age: 66.3±9.3 years; height: 163.8±8.0 cm;
weight: 74.3±15.2 kg), 17 with H-Y = 2 (age: 72.9±8.4 years; height: 164.4±9.6 cm; weight: 72.9
±12.7 kg), 27 with H-Y = 3 (age: 68.9±8.9 years; height: 162.9±8.7 cm; weight: 75.0±11.4 kg),
and 12 with H-Y = 4 (age: 73.8±7.2 years; height: 162.3±7.9 cm; weight: 71.1±10.8 kg)). The
inclusion criteria were: a diagnosis of idiopathic PD according to the UK Brain Bank Diagnos-
tic Criteria [44], H-Y stages 1–4, stable drug program (taking current medication for at least 2
weeks) and the ability to walk independently on at least the laboratory pathway without show-
ing freezing of gait. The exclusion criteria were: cognitive deficit (defined as scores <24 on the
Mini-Mental State Examination, moderate or severe depression (defined as scores �20 on the
Beck Depression Inventory, and presence of orthopedic and/or other gait-influencing condi-
tions such as arthrosis or total hip joint replacement. Medication was kept constant through-
out the trial period and all interventions were performed at the same time of day for each
patient during the “ON phase.” The participants were asked to maintain their usual activity
levels and current medication dosage when not in the laboratory. The assessments for both
clinical and instrumental evaluations were not involved in the treatment of the patients and
were blinded to the time of the evaluation. The patients were on oral levodopa (16 patients),
dopamine agonists (27 patients), or both (33 patients) and were recorded to be in the “ON
phase.” The levodopa equivalent dose was 571.92 ± 317.2 mg [45].
Sixty-seven healthy subjects (HS) were enrolled as the healthy control group (age, 69.84
±8.79 years).
All participants provided written informed consent before taking part in the study, which
was approved by a local ethics committee (Sapienza University of Rome, Policlinico Umberto
I, UP 00988_2020) and complied with the principles of the Declaration of Helsinki.

2.2 Gait analysis

An optoelectronic motion analysis system (SMART-DX 6000 System, BTS, Milan, Italy) con-
sisting of six infrared cameras (sample frequency, 340 Hz) was used to detect the movement of
twenty-two passive spherical markers covered with reflective aluminum powder (15 mm in

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 3 / 16

PLOS ONE Neural Networks for Parkinson classification

diameter) placed over prominent bony landmarks, according to the International Society of
Biomechanics recommendations [46] and Davis’s protocol [47]. Using double-adhesive tape,
the markers were placed over the cutaneous projections of the spinous processes of the seventh
cervical vertebra and sacrum and bilaterally over the acromion, anterior superior iliac spine,
great trochanter, lateral femoral condyle, fibula head, lateral malleolus, and metatarsal head. In
addition to markers directly applied to the skin, sticks or wands, varying in length from 7 to 10
cm, placed at 1/3 of the length of the body segment (femur and leg) were also used [47].

2.3 Experimental procedure

The patients and controls were asked to walk barefoot at a comfortable, self-selected speed
along a walkway approximately 10 m in length while looking forward. We provided only gen-
eral and qualitative instructions because we were interested in natural locomotion. Before the
recording session, the subjects practiced for a few minutes to familiarize themselves with the
procedure. We instructed the HCs to also walk at low speeds to compare the parameters
between groups without potential velocity bias [48]. At least ten trials were recorded for
PwPD. At least ten trials at a self-selected speed and ten trials at a slow speed were recorded for
HCs. Groups of three trials were separated by 1-min rest periods to avoid fatigue. To avoid
potential velocity bias, gait speed was matched between groups as follows: for each healthy sub-
ject, we considered only those trials in which gait speed fell within the range identified by
PwPD mean gait speed ± standard deviation.

2.4 Data analysis

After each acquisition performed by Smart Capture (BTS, Milan, Italy), three-dimensional
marker trajectories were reconstructed using a frame-by-frame tracking system (SMART
Tracker, BTS, Milan, Italy). Then, the data were processed using SMART Analyzer (BTS,
Milan, Italy) and MATLAB (version 7.10.0, MathWorks, Natick, MA, USA) software.
In this study, heel strike and toe-off events were determined by maximum and minimum
limb angle excursions. The limb angle was calculated as the angle between the vertical axis
from the greater trochanter and a vector drawn from the greater trochanter to the lateral mal-
leolus projected on the sagittal plane: a 0˚ limb angle meant that the leg was positioned verti-
cally under the body; positive angles denoted flexion (i.e., limb positioned in front of the
vertical axis) and negative angles denoted extension (i.e., limb positioned behind the vertical
axis) [49]. After this preprocessing procedure, the time-distance and kinematic parameters
were evaluated and the kinematic data were normalized to the duration of the gait cycle
(defined as the interval between two successive foot contacts of the same leg) and interpolated
to 101 samples using a polynomial procedure.
2.4.1 Time-distance parameters. The following time-distance gait parameters were calcu-
lated for each subject and for each stride: walking speed (m/s), cadence (step/s), step width
(m), step length (m) (defined as the distance from the heel strike of a limb and the subsequent
heel strike of the other limb), stance, swing, and double support phase durations (expressed as
percentages of the gait cycle duration). Step length and step width were normalized to the limb
length of each subject. For each subject, the average value of each gait feature was calculated.
2.4.2 Kinematic data. The anatomical joint angles of the hip, knee, ankle, trunk, and pel-
vis (frontal, sagittal, and transverse planes) and the corresponding ranges of motion (RoMs) of
the joints (defined as the differences between the maximum and minimum values during the
gait cycle) were computed. For each subject, the average value of each RoM was calculated.
2.4.3 ANN approach for the diagnosis and staging of the gait deficit in PwPD. HS vs
PwPD classification. A principal component analysis (PCA), using a threshold of 98% on the

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 4 / 16

PLOS ONE Neural Networks for Parkinson classification

cumulative variance was used to define a subset of features starting from all time-distance and
kinematic HS and PwPD features [50].
Then, an artificial neural network (ANN) approach was used for diagnosis of Parkinson
disease using the features selected by PCA.
We trained different topologies of feedforward networks with different numbers of hidden
layers (HL) and different numbers of neurons (N) in each HL. The number of HL varied in the
range of 1–3, while the number N in each HL varied based on the number of N in the first hid-
den layer (NL1), N was set to two different values (20 and 50, respectively), and the number of
nodes in the other HL (when defined) was 1/2 and 1/3 of N for the second (NL2) and third
(NL3) HLs, respectively. Thus, the combination of L layers and N nodes in the first HL hidden
layer led to the six different network architectures. The output set consisted of an orthogonal
coding of the two values (HS vs PwPD): OUT1 = [1 0] and OUT2 = [0 1] (one-hot classifica-
tion coding scheme).
Networks were trained with a supervised approach using the Levenberg-Marquardt back-
propagation algorithm, stopping when at least one of the following conditions was met: 1000
iterations, 10−6 mean square error, or six consecutive fails on the validation set [17]. To verify
the repeatability of our results, each of the six network topologies was trained ten times by
using a random 10% of samples as the validation set and a random 10% as the testing set. For
each trained network, a confusion matrix was calculated based on the real value (HS or PwPD)
and the one estimated on the randomly extracted testing set.
The mean 2×2 confusion matrix was then obtained by averaging the confusion matrixes of
the trained ANNs. A performance parameter (P) was calculated as the mean (%) of the ele-
ments on the diagonal of the mean confusion matrix, where 100% indicates the absence of mis-
classifications [17]. Furthermore, the sensitivity and specificity of each group were calculated.
The entire system is schematically described in Fig 1.
Then, to reduce the features, we also used subsets of features from the selected features with
PCA and, for each subset, we trained the six ANNs ten times to evaluate the confusion matrix
and performance. We started with subsets of all combination of two features and continued
until we identified a subset whose performance was no different from that of the set with all
features selected with PCA (Fig 1).
PD staging classification. An ANN approach was also used to stages the gait deficits in PD in
terms of the H-Y scale using the features selected by PCA (2.4.3.2) and walking speed.
We trained the six different topologies of feedforward networks as described in section
2.4.3.1. The output set consisted of an orthogonal coding of four values of H-Y: OUT1 = [1 0 0
0], OUT2 = [0 1 0 0], OUT3 = [0 0 1 0], and OUT4 = [0 0 0 1] (one-hot classification coding
scheme). For each trained network, a confusion matrix was calculated based on the real H-Y
value and the one estimated on the randomly extracted testing set. The mean 4×4 confusion
matrix was then obtained by averaging the confusion matrices of the trained ANNs and the P
was calculated [17].
As described in section 2.4.3.1, to reduce the features, we also used subsets of features from
among the features selected by PCA and, for each subset, we trained the six ANNs ten times,
evaluating the confusion matrix and performance. We started with all combination of two fea-
tures subsets and continued until we found a subset whose performance was no different from
that of the set with all features selected by PCA (Fig 1).

2.5. Statistical analysis

The Shapiro–Wilk test for normal distribution was preliminarily executed on all gait parame-
ters. Unpaired two-sample t-tests or Mann–Whitney tests (two-tailed) were used for

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 5 / 16

PLOS ONE Neural Networks for Parkinson classification

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 6 / 16

PLOS ONE Neural Networks for Parkinson classification

Fig 1. Description of experimental set-up and methodological approach. A schematic description of the walking and artificial neural network method used to
map time-distance and kinematic features on the H&Y (1, 2, 3 and 4) levels.
https://doi.org/10.1371/journal.pone.0244396.g001

assessment of between-group differences in the time-distance parameters and joint kinematics

values. Then, we performed a two-way analysis of variance (ANOVA) test with L and N as fac-
tors to determine the possible significant effects on ANN performance, sensitivity and specific-
ity caused by the listed factor. Separate ANOVAs test were performed for performance,
sensitivity and specificity. Post-hoc analysis with Bonferroni’s corrections was performed
when significant differences were observed in the ANOVA results. P values < 0.05 were con-
sidered statistically significant.
Statistical analysis was performed to check if the results of the performance obtained using
all features of PCA differed significantly from those obtained considering the subsets of two,
three, or four features.
As a confirmative analysis, independent samples t-tests and univariate ANOVA with Bon-
ferroni post-hoc analysis were performed to test the ability of the identified minimum sets of
gait parameters to differentiate between PwPD and HS, and PwPD across the H-Y stages,
respectively. Receiver operating characteristic (ROC) curves were plotted to assess the discrim-
inative ability of the identified minimum sets of gait parameters in differentiating PwPD from
HS and PwPD across the H-Y stages. Area under the curve (AUC), sensitivity and specificity,
and positive (LR+) and negative (LR-) likelihood ratios were calculated. The optimal cutoff
points (OCP) for the cumulative indices of the combinations of gait parameters included in
the identified sets were calculated as the point of the ROC curve where the sum of sensitivity
and specificity was highest. Post-test probabilities were inferred by transforming LRs into odds
ratios using the Fagan nomogram [51, 52].

3. Results
3.1 Time-distance and joint kinematics parameters
The values of the time-distance and joint kinematic parameters are reported in Table 1 for
both groups. Compared with HCs, PwPD showed significantly lower step length; stride length;
hip, knee, and ankle RoMs; trunk flexion-extension; trunk rotation and pelvis rotation values;
and higher cadence.

3.2 HS vs. PwPD classification

The PCA showed that a set of nine features (stance duration, swing duration, step length,
ankle, knee, hip, trunk rotation, pelvis rotation, and pelvis tilt RoM) expressed 98% of the
cumulative variance of the data (Fig 2A).
The ANNs analysis revealed a mean performance range of 82–98% (Fig 2A), a mean sensi-
tivity range of 81–89% (Table 2), and a mean specificity range of 81–89% (Table 2) for all fea-
tures detected by PCA for each number of neurons and HL. The number of neurons did not
affect the performance (F1 = 0.15, p = 0.699), sensitivity (F1 = 0.5, p = 0.482), or specificity (F1
= 0.5, p = 0.482) in two-way ANOVA. Furthermore, the HL did not affect the performance (F2
= 0.48, p = 0.621), sensitivity (F2 = 0.32, p = 0.724), or specificity (F2 = 0.48, p = 0.621) in two-
way ANOVA.
By analyzing all the possible combinations, we found that one combination of two features
was the minimum set of gait parameters able to distinguish PwPD from controls (knee RoM,
trunk rotation RoM) and whose performance (Fig 2A) did not significantly differ (p>0.05)
from that of all PCA features.

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 7 / 16

PLOS ONE Neural Networks for Parkinson classification

Table 1. Kinematic parameters.

Parameters PwPD HS
Spatio-temporal parameters Gait speed (km/h) 2.87±1.07 3.18±0.91
Stance duration (% gait cycle) 65.66±3.49 65.54±3.13
Swing duration (% gait cycle) 34.34±3.49 34.46±3.13
1st double support (% gait cycle) 15.74±3.68 15.40±3.16
2nd double support (% gait cycle) 15.48±3.42 15.50±3.20
Step length (% limb length) 0.68±0.14� 0.77±0.10
Step width (% limb length) 0.30±0.05 0.28±0.05
�
Stride length (% limb length) 1.19±0.32 1.43±0.23
Cadence 0.89±0.22� 0.77±0.15
Range of motion RoM Hip 32.97±11.89� 37.38±5.23
RoM Knee 52.28±9.39� 60.09±10.62
RoM Ankle 24.19±7.26� 32.00±12.05
Trunk lateral bending [˚] 3.31±2.05 4.05±1.66
Trunk flexion-extension [˚] 2.84±1.10� 5.99±2.45
Trunk rotation [˚] 8.64±5.05 � 14.36±11.12
Pelvis obliquity [˚] 4.82±3.37 5.77±2.06
Pelvis rotation [˚] 11.20±8.55� 16.48±13.26
Pelvis tilt [˚] 67.78±21.98� 84.25±24.43

Mean ± SD of time-distance and kinetic parameters in patients with Parkinson disease (PwPD) patients and healthy
subjects (HS).
�
p-value<0.05.

https://doi.org/10.1371/journal.pone.0244396.t001

The results of the independent sample t–test and ROC curve analysis confirmed that the
combination of knee and trunk rotation RoM values could significantly differentiate between
PwPD and HS (t-statistic = −5.34, p<0.00) and to have good discriminative ability (AUC =
0.77). The numerical sum of knee and trunk rotation RoMs � 66.23 was able to identify PwPD
from HS with a 75% probability (Table 3).

3.3 PwPD staging classification

The ANNs analysis revealed a mean performance range of 66.16–77.2% (Fig 2B), mean sensi-
bility range of 66–77% (Table 2) and mean specificity range of 85–91% (Table 2) for all features
detected by PCA for each number of neurons and HLs.
The number of neurons had no effect on the performance (F1 = 2.82, p = 0.099), sensitivity
(F1 = 0.41, p = 0.522), and specificity (F1 = 0.59, p = 0.448) in two-way ANOVA. Furthermore,
the HL had no effect on the performances (F2 = 0.66, p = 0.522), sensitivity (F2 = 0.7,
p = 0.499), and specificity (F2 = 0.29, p = 0.749) in two-way ANOVA.
By analyzing all the possible combinations, we found that two combinations of four features
(walking speed and hip, knee, and ankle RoMs; walking speed and hip, knee, and trunk rota-
tion RoMs) were the minimum set of gait parameters able to distinguish H-Y stage gait pat-
terns from one another and whose performances (Fig 2B) did not differ significantly (p>0.05)
from that of all PCA features. All combinations of two or three features showed a significant
difference (p<0.05) from that of all PCA features.
The numerical sums of speed, hip RoM, knee RoM, and ankle RoM (SET1) and speed, hip
RoM, knee RoM, and trunk rotation RoM (SET2) were able to differentiate PwPD according

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 8 / 16

PLOS ONE Neural Networks for Parkinson classification

Fig 2. Accuracy of artificial neural networks and the best mean confusion matrix. For diagnosis (A) and staging (B), in the first row the accuracy of artificial neural
networks and in the second row the best mean confusion matrixes considering all PCA features as INPUT (a1 and b1) and subset of 2 features (knee RoM and trunk
rotation RoM (a2)) and subsets of 4 features (walking speed, hip, knee and ankle RoMs (b2); walking speed, hip, knee and trunk rotation RoMs (b3)). Six different
architectures of neural networks were represented by varying the numbers of hidden layers (1, 2, or 3) and the numbers of neurons in each hidden layer based on the
numbers of nodes N in the first hidden layer.
https://doi.org/10.1371/journal.pone.0244396.g002

to H-Y stage (F = 7.59, p<0.00 and F = 9.27, p<0.00, respectively). Post-hoc analysis revealed
that SET1 was significantly different between H-Y stages 1 and 4 (p<0.00) and stages 3 and 4
(p = 0.02), while SET2 was able to differentiate between H-Y stages 1 and 2 (p = 0.03), 1 and 3
(p = 0.03), 1 and 4 (p<0.00), and 3 and 4 (p = 0.03). The AUCs, OCPs, sensitivity, specificity,
LRs, and post-test probabilities of each set to discriminate PwPD across the H-Y stages are
summarized in Table 3.
Briefly, PwPD at H-Y stage 1 were identified by cumulative SET1 values � 111.71 and SET2
values � 105.42, PwPD at H-Y stage 2 by cumulative SET2 values � 103.91 and � 105.42,
PwPD at H-Y stage 3 by SET2 values � 81.35 and � 103.91, and PwPD at H-Y stage 4 by SET1
values � 99.28 and SET2 values � 81.35.

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 9 / 16

PLOS ONE Neural Networks for Parkinson classification

Table 2. Sensitivity and specificity.

HS vs PwPD classification. PwPD staging classification

L N HS PwPD Mean H&Y = 1 H&Y = 2 H&Y = 3 H&Y = 4 Mean
sensibility 1 20 93.2 ± 2.03 78.32 ± 11.2 85.76 ± 5.66 87.88 ± 5.52 61.27 ± 15.99 69.09 ± 8.41 61.05 ± 17.4 69.83 ± 7.36
[%] 2 20 91.76 ± 3.71 81.89 ± 8.24 86.83 ± 4.77 89.81 ± 5.37 66.18 ± 15.24 74.29 ± 7.08 74.47 ± 18.11 76.19 ± 8.31
3 20 91.89 ± 4.53 73.58 ± 27.59 82.74 ± 12.54 87.69 ± 9.34 52.36 ± 21.66 70.65 ± 12.61 53.95 ± 21.1 66.16 ± 13.1
1 50 94.19 ± 1.48 80.63 ± 7.33 87.41 ± 3.27 90.77 ± 4.23 69.27 ± 18.31 75.84 ± 7.27 72.89 ± 13.19 77.2 ± 7.75
2 50 92.97 ± 1.69 70 ± 22.21 81.49 ± 10.85 86.92 ± 11.13 66 ± 20.21 71.04 ± 12.78 71.58 ± 22.32 73.89 ± 11.23
3 50 92.79 ± 2.11 84.84 ± 8.04 88.82 ± 4.21 89.23 ± 7.03 67.64 ± 24.87 72.34 ± 22.87 73.95 ± 24.79 75.79 ± 16.41
specificity 1 20 78.32 ± 11.2 93.2 ± 2.03 85.76 ± 5.66 72.32 ± 7.03 92.12 ± 5.05 90.66 ± 10.85 96.86 ± 5.14 87.99 ± 3.58
[%] 2 20 81.89 ± 8.24 91.76 ± 3.71 86.83 ± 4.77 79.1 ± 10.47 95.12 ± 5.21 91.23 ± 5.7 97.91 ± 1.84 90.84 ± 3.76
3 20 73.58 ± 27.59 91.89 ± 4.53 82.74 ± 12.54 67.5 ± 17.57 91.71 ± 5.63 87.11 ± 11.05 97.5 ± 3.04 85.95 ± 7.11
1 50 80.63 ± 7.33 94.19 ± 1.48 87.41 ± 3.27 80.68 ± 9.94 95.03 ± 3.76 90.56 ± 4.89 98.53 ± 0.73 91.2 ± 3.49
2 50 70 ± 22.21 92.97 ± 1.69 81.49 ± 10.85 75.07 ± 17.92 95.32 ± 2.73 92.24 ± 7.43 98.39 ± 1.43 90.26 ± 4.24
3 50 84.84 ± 8.04 92.79 ± 2.11 88.82 ± 4.21 76.91 ± 25.38 95.38 ± 3.82 94.72 ± 4.78 98.2 ± 1.26 91.3 ± 5.38

Sensitivity and Specificity of set with all PCA features for HS vs PwPD classification and for PwPD staging classification.

https://doi.org/10.1371/journal.pone.0244396.t002

4. Discussion
Our study was conducted to address the enhancement of the diagnosis and staging of the Par-
kinsonian gait using an automated machine-learning technique (i.e., ANNs). We found that
the accuracy of ANNs obtained with all PCA features did not differ from that obtained with
the combination of two features (Fig 2, Table 2); namely, knee and trunk rotation RoMs, sug-
gesting the consideration of only a minimum set of two features to distinguish PwPD from
HCs. In the confirmatory analysis, the combination of both knee and trunk rotation RoMs, as
the numerical sum of knee and trunk rotation angles, showed a good ability to discriminate
PwPD from HCs, with a cumulative threshold value of � 66.23˚ (Table 3). Knee and trunk
rotation RoM abnormalities characterize the gait pattern of PwPD, as found in the current
study (Table 1) and in previous studies [53–58], which also revealed a series of other kinematic
gait abnormalities. Our findings indicate an ANN algorithm resulted in a drastic reduction in
the amount of redundant information, allowing a focus on a few meaningful features that

Table 3. Ability to discriminate between PwPD and HS and between disability levels.
HS vs PwPD HY 2 vs 1 HY 3 vs1 HY 4 vs 1 HY 4 vs 3
SET SET1 SET2 SET1 SET2 SET1 SET2 SET1 SET2
AUC 0.77 n.a. 0.73 n.a. 0.76 0.88 0.88 0.77 0.78
OCP � 66.23 n.a. � 105.42 n.a. � 103.91 �111.71 �98.09 �99.28 �81.35
Se (%) 72.37 n.a. 76.47 n.a. 81.48 91.67 75 66.67 58.35
Sp (%) 73.13 n.a. 60 n.a. 70 80 92 88.89 96.30
LR+ 2.69 n.a. 1.91 n.a. 2.72 4.58 3.66 6 15.75
LR- 0.38 n.a. 0.39 n.a. 0.26 0.10 0.11 0.37 0.43
+PTP (%) 75 n.a. 62 n.a. 79 73 69 73 88
-PTP (%) 30 n.a. 25 n.a. 26 6 6 14 16

HS = Healthy Subjects; PwPD = people with Parkinson’s Disease; HY = Hoehn & Yahr disability stage; SET = Knee RoM+Trunk rotation RoM; SET1 = combination of
gait speed, hip, knee and ankle Roms values; SET2 = combination of gait speed, hip, knee and trunk rotation RoMs values; AUC = area under the curve; OCP = optimal
cutoff point; Se = sensitivity; Sp = specificity; LR+ = positive likelihood ratio; LR- = negative likelihood ratio; +PTP = positive post-test probability: the probability to
identify a true positive at OCP; -PTP = negative post-test probability: the probability to identify a false negative at OCP.

https://doi.org/10.1371/journal.pone.0244396.t003

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 10 / 16

PLOS ONE Neural Networks for Parkinson classification

could be used to diagnose and monitor gait function in PwPD. Both the knee and trunk rota-
tion RoMs showed highly discriminatory features characterizing Parkinsonian gait. With
regard to the classifier knee joint RoM, our findings are in line with that in study by Caramia
et al. [28] using an inertial measurement units-based classification revealed that the knee joint
played a major role among the lower limb joints in the assessment of gait in PwPD. PwPD
often show lower limb joint rigidity and bradykinesia [59, 60], deficits in knee joint muscle
strength and endurance [61, 62], abnormal knee-bent posture [63, 64], and reduced knee
motion during walking [65]. Consequently, knee motor control can be impaired in its role of
producing an adequate extensor moment [66] and absorbing foot-ground impact [67] in the
loading response, recovering and storing energy in the mid-stance [68], and allowing limb
progression in the swing phase.
The flexed trunk posture has been reported as a classifier of trunk rotation in PwPD [61–
71]. Although the use of the terms camptocormia and Pisa syndrome are restricted to the
patients with extreme trunk flexion (>60˚) or lateral bending (>10˚–15˚) [72], who were not
included in our study, a certain degree of trunk posture abnormalities is present in most
PwPD [73]. In recent decades, a series of quantitative motion analysis studies have also
revealed restrained movement of the trunk during several locomotion tasks [74, 75]. Reduced
trunk motion combined with postural abnormalities can greatly alter the role of the spine in
balance maintenance [6, 68, 75–78], ultimately predisposing patients to falls [71]. Notably, a
recent study showed that trunk rotation is a predictor of gait recovery after rehabilitation [57],
suggesting that rehabilitation should focus on recovering trunk control [79, 80] to improve
both gait and balance [56–58]. Altogether, these findings underline the importance of consid-
ering trunk kinematic abnormalities as an integral part of the gait deficit in PwPD.
When classifying the gait deficit of PwPD according to the H-Y staging system, we found
that two combinations of four features (SET1: walking speed and hip, knee, and ankle RoMs;
SET2: walking speed and hip, knee, and trunk rotation RoMs, respectively,) were the mini-
mum set of gait parameters that were needed to distinguish the gait patterns of PwPD at differ-
ent H-Y stages (Fig 2, Table 2). Their performance did not significantly differ from that of all
PCA features (Fig 2, Table 2). These findings emphasize the usefulness of reduced lower limb
joint RoMs as the main classifiers in addition to trunk motion and slow speed, all of which rep-
resent sensitive diagnostic features of gait function decline in PwPD. Since our algorithm
included 18 gait kinematic variable classifiers, the possibility of focusing on two sets of few gait
variables (Fig 2, Table 3) drastically simplifies the diagnosis of gait deficit staging and, thus,
might assist physicians in monitoring gait function in PwPD.
Our results should be viewed considering the study limitations and in comparison, with
three recent studies on gait classification applying machine-learning techniques in
PwPD28,32,35. Patients undergoing clinical gait analysis often walk slower than healthy people.
Since many gait variables are speed-dependent [79–85], thus, in our study, the control and
patient groups were speed-matched to avoid bias due to gait speed. This procedure allowed us
to focus on the most discriminative parameters, irrespective of the gait speed. Previous studies
on gait deficit classification [28, 31, 34] did not control for gait speed when comparing patients
with healthy subjects, and thus, may have given importance to some speed-related classifiers.
However, when classifying the gait deficit of PwPD according to the H-Y staging system, we
could not control for gait speed. One could argue that joint RoMs might be influenced by the
gait speed itself more than the disease severity. However, a robust body of evidence has shown
a worsening of lower limb rigidity, bradykinesia [86], and joint RoMs [87] in the advanced
stages of the disease, suggesting that reduced gait speed is a consequence of limb rigidity/bra-
dykinesia and not vice versa.

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 11 / 16

PLOS ONE Neural Networks for Parkinson classification

The present results should inspire future research to test the utility of gait classifiers for the
differential diagnosis of different neurological gait disorders. Furthermore, longitudinal stud-
ies are needed to verify whether the set of gait classifiers represent sensitive tools to monitor
the effect of pharmacological and rehabilitation treatments that can modify disease
progression.

4.1. Limitations
The lack of a comparison of our results with gold standard clinical measures for gait assess-
ment in PwPD as well as with other linear or non-linear statistical models could represent a
limitation of this study. Notwithstanding, the probabilities to discriminate PwPD from healthy
subjects, and PwPD across the disease progression stages, through the gait parameters reported
in our model, is similar to that reported in other studies on PwPD and other neurological dis-
orders [43, 88, 89]. However, there is significant room for improving the accuracy of our
model. For instance, other biomarkers of gait impairment, (i.e. arms oscillation, gait variabil-
ity, kinetic and muscle activation variables) or other dimensions of disability (i.e. cognitive
impairment, self-reported quality of life), not included in our analysis, could enhance the accu-
racy in identifying disability status through gait analysis.

Author Contributions
Conceptualization: Tiwana Varrecchia, Stefano Filippo Castiglia, Mariano Serrao.
Data curation: Tiwana Varrecchia, Stefano Filippo Castiglia.
Formal analysis: Tiwana Varrecchia.
Investigation: Tiwana Varrecchia, Alberto Ranavolo, Carmela Conte, Antonella Tatarelli,
Gianluca Coppola, Cherubino Di Lorenzo, Mariano Serrao.
Methodology: Tiwana Varrecchia.
Software: Tiwana Varrecchia.
Supervision: Francesco Draicchio, Francesco Pierelli, Mariano Serrao.
Validation: Tiwana Varrecchia, Mariano Serrao.
Visualization: Tiwana Varrecchia, Mariano Serrao.
Writing – original draft: Tiwana Varrecchia, Stefano Filippo Castiglia, Alberto Ranavolo,
Mariano Serrao.
Writing – review & editing: Tiwana Varrecchia, Stefano Filippo Castiglia, Alberto Ranavolo,
Carmela Conte, Antonella Tatarelli, Gianluca Coppola, Cherubino Di Lorenzo, Francesco
Draicchio, Francesco Pierelli, Mariano Serrao.

References
1. Bhidayasiri R and Martinez-Martin P. Clinical Assessments in Parkinson’s Disease: Scales and Monitor-
ing. Int Rev Neurobiol 2017; 132: 129–182. https://doi.org/10.1016/bs.irn.2017.01.001 PMID:
28554406
2. Jankovic J. Parkinson’s disease: clinical features and diagnosis. J Neurol Neurosurg Psychiatry 2008;
79(4): 368–376. https://doi.org/10.1136/jnnp.2007.131045 PMID: 18344392
3. Serrao M, Chini G, Bergantino M, et al. Identification of specific gait patterns in patients with cerebellar
ataxia, spastic paraplegia, and Parkinson’s disease: A non-hierarchical cluster analysis. Hum Mov Sci
2018; 57: 267–279. https://doi.org/10.1016/j.humov.2017.09.005 PMID: 28967438

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 12 / 16

PLOS ONE Neural Networks for Parkinson classification

4. Martinez-Martin P, Rodriguez-Blazquez C, Forjaz MJ, Kurtis MM. Impact of pharmacotherapy on quality

of life in patients with Parkinson’s disease. CNS Drugs 2015; 29, 397–413. https://doi.org/10.1007/
s40263-015-0247-x PMID: 25968563
5. Pickering RM, Grimbergen YA, Rigney U, et al. A meta-analysis of six prospective studies of falling in
Parkinson’s disease. Movement Disorders 2007; 22: 1892–1900. https://doi.org/10.1002/mds.21598
PMID: 17588236
6. Morrone M, Miccinilli S, Bravi M, et al. Perceptive rehabilitation and trunk posture alignment in patients
with Parkinson disease: a single blind randomized controlled trial. Eur J Phys Rehabil Med 2016; 52(6):
799–809. PMID: 27171537
7. Viitanen M, Mortimer JA, Webster DD. Association between presenting motor symptoms and the risk of
cognitive impairment in Parkinson’s disease. J Neurol Neurosurg Psychiatry 1994; 57(10): 1203–1207.
https://doi.org/10.1136/jnnp.57.10.1203 PMID: 7931381
8. Pons JL, Moreno JC, Torricelli D, Taylor JS. Principles of human locomotion: a review. Conf Proc IEEE
Eng Med Biol Soc 2013; 2013:6941–4. https://doi.org/10.1109/EMBC.2013.6611154 PMID: 24111341
9. McCrone P, Allcock LM, Burn DJ. Predicting the cost of Parkinson’s disease. Mov Disord 2007; 22:
804–12. https://doi.org/10.1002/mds.21360 PMID: 17290462
10. Baker R, Esquenazi A, Benedetti MG, Desloovere K. Gait analysis: clinical facts. Eur J Phys Rehabil
Med 2016; 52(4): 560–574. PMID: 27618499
11. Ferrarin M, Bovi G, Rabuffetti M, et al. Gait pattern classification in children with Charcot-Marie-Tooth
disease type 1A. Gait Posture 2012; 35: 131–137. https://doi.org/10.1016/j.gaitpost.2011.08.023
PMID: 21944474
12. Kyeong S, Kim SM, Jung S, Kim DH. Gait pattern analysis and clinical subgroup identification: a retro-
spective observational study. Medicine (Baltimore): 2020; 99(15): 19555. https://doi.org/10.1097/MD.
0000000000019555 PMID: 32282704
13. Sawacha Z, Sartor CD, Yi LC, Guiotto A, Spolaor F, Sacco ICN. Clustering classification of diabetic
walking abnormalities: a new approach taking into account intralimb coordination patterns. Gait Posture
2020; 79: 33–40. https://doi.org/10.1016/j.gaitpost.2020.03.016 PMID: 32334348
14. Serrao M, Rinaldi M, Ranavolo A, et al. Gait patterns in patients with hereditary spastic paraparesis.
PLoS ONE 2016; 11: 1–16. https://doi.org/10.1371/journal.pone.0164623 PMID: 27732632
15. Zurada J, Karwowski W, Marras W. A neural network-based system for classification of industrial jobs
with respect to risk of low back disorders due to workplace design. Appl Ergon 1997; 28(1): 49–58.
https://doi.org/10.1016/s0003-6870(96)00034-8 PMID: 9414340
16. Chen CL, Kaber DB, Dempsey PG. Using feedforward neural networks and forward selection of input
variables for an ergonomics data classification problem. Hum Factors Ergon Manuf 2004; 14(1): 31–
49. http://doi:10.1002/hfm.10052.
17. Varrecchia T, De Marchis C, Rinaldi M, et al. Lifting activity assessment using surface electromyo-
graphic features and neural networks. Int J Ind Ergon 2018; 66: 1–9.
18. Varrecchia T, De Marchis C, Draicchio F, Schmid M, Conforto S, Ranavolo A. Lifting Activity Assess-
ment Using Kinematic Features and Neural Networks Appl Sci 2020; 10(6): 1989. https://doi.org/10.
3390/app10061989
19. Ahmed F. Artificial neural networks for diagnosis and survival prediction in colon cancer. Mol Cancer
2005; 4: 29. https://doi.org/10.1186/1476-4598-4-29 PMID: 16083507
20. Spelt L, Andersson B, Nilsson J, Andersson R. Prognostic models for outcome following liver resection
for colorectal cancer metastases: A systematic review. Eur J Surg Oncol 2012; 38: 16–24. https://doi.
org/10.1016/j.ejso.2011.10.013 PMID: 22079259
21. Mortazavi D, Kouzani A, Soltanian-Zadeh H. Segmentation of multiple sclerosis lesions in MR images:
a review. Neuroradiology. 2012a; 54: 299–320. https://doi.org/10.1007/s00234-011-0886-7 PMID:
21584674
22. Mortazavi D, Kouzani AZ, Soltanian-Zadeh H. Segmentation of multiple sclerosis lesions in MR images:
a review. Neuroradiology. 2012b; 54: 299–320. https://doi.org/10.1007/s00234-011-0886-7 PMID:
21584674
23. Bartosch-Härlid A, Andersson B, Aho U, Nilsson J, Andersson R. Artificial neural networks in pancreatic
disease. Br J Surg 2008; 95: 817–826. https://doi.org/10.1002/bjs.6239 PMID: 18551536
24. Siristatidis C, Chrelias C, Pouliakis A, Katsimanis E, Kassanos D. Artificial neural networks in gyneaco-
logical diseases: Current and potential future applications. Med Sci Monit 2010; 16: 231–236.
25. Shankaracharya DO, Samanta S, Vidyarthi A. Computational intelligence in early diabetes diagnosis: a
review. Rev Diabet Stud 2010; 7: 252–262. https://doi.org/10.1900/RDS.2010.7.252 PMID: 21713313

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 13 / 16

PLOS ONE Neural Networks for Parkinson classification

26. Zhang S, Poon SK, Vuong K, Sneddon A, Loy CT. A Deep Learning-Based Approach for Gait Analysis
in Huntington Disease. Stud Health Technol Inform 2019; 64: 477–481. https://doi.org/10.3233/
SHTI190267 PMID: 31437969
27. Rovini E, Maremmani C, Cavallo F. How wearable sensors can support parkinson’s disease diagnosis
and treatment: A systematic review. Front Neurosci. 2017; 11: 555. https://doi.org/10.3389/fnins.2017.
00555 PMID: 29056899
28. Caramia C, Torricelli D, Schmid M, et al. IMU-Based Classification of Parkinson’s Disease From Gait: A
Sensitivity Analysis on Sensor Location and Feature Selection. IEEE J Biomed Health Inform. 2018
Nov; 22(6): 1765–1774. https://doi.org/10.1109/JBHI.2018.2865218 PMID: 30106745
29. Pickle NT, Shearin SM, Fey NP. A machine learning approach to targeted balance rehabilitation in peo-
ple with Parkinson’s disease using a sparse sensor set. Conf Proc IEEE Eng Med Biol Soc. 2018;
1202–1205. https://doi.org/10.1109/EMBC.2018.8512530 PMID: 30440605
30. Benmalek E, Elmhamdi J, Jilbab A. Multiclass classification of parkinson’s disease using different clas-
sifiers and llbfs feature selection algorithm. Int J Speech Technol. 2017; 20(1): 179–84. https://doi.org/
10.1007/s10772-017-9401-9.
31. Buongiorno D, Bortone I, Cascarano GD, Trotta GF, Brunetti A, Bevilacqua V. A low-cost vision system
based on the analysis of motor features for recognition and severity rating of Parkinson’s Disease. BMC
Med Inform Decis Mak. 2019; 19(9): 243. https://doi.org/10.1186/s12911-019-0987-5 PMID: 31830986
32. Goetz CG, Tilley BC, Shaftman SR, et al. Movement disorder society-sponsored revision of the unified
parkinson’s disease rating scale (mds-updrs): scale presentation and clinimetric testing results. Mov
Disord Off J Mov Disord Soc. 2008; 23(15): 2129–70. https://doi.org/10.1002/mds.22340 PMID:
19025984
33. Hoehn MM, and Yahr MD. Parkinsonism onset, progression, and mortality. Neurology 1967; 17(5):
427–427. https://doi.org/10.1212/wnl.17.5.427 PMID: 6067254
34. Cavallo F, Moschetti A, Esposito D, Maremmani C, Rovini E. Upper limb motor pre-clinical assessment
in parkinson’s disease using machine learning. Parkinsonism Relat Disord 2019. https://doi.org/10.
1016/j.parkreldis.2019.02.028 PMID: 30826265
35. Goetz CG, Poewe W, Rascol O, et al. Movement Disorder Society Task Force report on the Hoehn and
Yahr staging system: status and recommendations. Mov Disord 2004; 19: 1020–1028. https://doi.org/
10.1002/mds.20213 PMID: 15372591
36. Staffen W, Mair A, Unterrainer J, Trinka E, Ladurner G. Measuring the progression of idiopathic Parkin-
son’s disease with 123I-b-CIT-SPECT. J Neur Trans 2000; 107: 543–552.
37. Chen XQ, Niu JP, Peng RQ, Song YH, Xu N, Zhang YW. The early diagnosis of Parkinson’s disease
through combined biomarkers. Acta Neurol Scand 2019; 140(4): 268–273. https://doi.org/10.1111/ane.
13140 PMID: 31190374
38. Jenkinson C, Fitzpatrick R, Peto V, Greenhall R, Hyman N. The Parkinson’s Disease Questionnaire
(PDQ39): development and validation of a Parkinson’s disease summary index score. Age Ageing
1997; 26: 353–357. https://doi.org/10.1093/ageing/26.5.353 PMID: 9351479
39. Reynolds NC, Montgomery GK. Factor analysis of Parkinson’s impairment. Arch Neurol 1987; 44:
1013–1016. https://doi.org/10.1001/archneur.1987.00520220019009 PMID: 3632372
40. Poewe W, Wenning GK. The natural history of Parkinson’s disease. Neurology 1996; 47: 146–152.
https://doi.org/10.1212/wnl.47.6_suppl_3.146s PMID: 8959983
41. http://www.appde.eu/european-physiotherapy-guidelines.asp
42. Buckley C, Alcock L, McArdle R, Rehman RZU, Del Din S, Mazzà C, Yarnall AJ, Rochester L. The Role
of Movement Analysis in Diagnosing and Monitoring Neurodegenerative Conditions: Insights from Gait
and Postural Control. Brain Sci. 2019; 9(2): 34. https://doi.org/10.3390/brainsci9020034 PMID:
30736374
43. Dinov ID, Heavner B, Tang M, Glusman G, Chard K, Darcy M, et al. Predictive Big Data Analytics: A
Study of Parkinson’s Disease Using Large, Complex, Heterogeneous, Incongruent, Multi-Source and
Incomplete Observations. PLoS One 2016; 11(8): e0157077. https://doi.org/10.1371/journal.pone.
0157077 PMID: 27494614
44. Hughes AJ, Daniel SE, Kilford L, Lees AJ. Accuracy of clinical diagnosis of idiopathic Parkinson’s dis-
ease: a clinico-pathological study of 100 cases. J Neurol Neurosurg Psychiatry 1992; 55: 181–4.
https://doi.org/10.1136/jnnp.55.3.181 PMID: 1564476
45. Tomlinson CL, Stowe R, Patel S, Rick C, Gray R, Clarke CE. Systematic review of levodopa dose equiv-
alency reporting in Parkinson’s disease. Mov Disord. 2010; 25(15): 2649–2653. https://doi.org/10.
1002/mds.23429 PMID: 21069833
46. Wu G, van der Helm FC, Veeger HE, et al. ISB recommendation on definitions of joint coordinate sys-
tems of various joints for the reporting of human joint motion–Part II: Shoulder, elbow, wrist and hand.

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 14 / 16

PLOS ONE Neural Networks for Parkinson classification

Journal of Biomechanics 2005; 38: 981–992. https://doi.org/10.1016/j.jbiomech.2004.05.042 PMID:

15844264
47. Davis RB, Ounpuu S, Tyburski D, Gage JR A gait analysis data collection and reduction technique.
Human Movement Science 1991; 10: 575–587.
48. Rinaldi M, Ranavolo A, Conforto S, et al. Increased lower limb muscle coactivation reduces gait perfor-
mance and increases metabolic cost in patients with hereditary spastic paraparesis. Clin Biomech (Bris-
tol, Avon) 2017; 48:63–72. https://doi.org/10.1016/j.clinbiomech.2017.07.013 PMID: 28779695
49. Borghese NA, Bianchi L, Lacquaniti F. Kinematic determinants of human locomotion. J Physiol 1996;
494(3): 863±879. https://doi.org/10.1113/jphysiol.1996.sp021539 PMID: 8865081
50. Jolliffe I T. Principal component analysis. Springer-Verlag, New York 1986.
51. Deeks JJ and Altman DG. Diagnostic tests 4: likelihood ratios. BMJ 2004; 329(7458): 168–9. https://
doi.org/10.1136/bmj.329.7458.168 PMID: 15258077
52. Wald NJ and Bestwick JP. The area under the ROC curve: is it a valid measure of screening perfor-
mance? J Med Screen 2014; 21(4): 220. https://doi.org/10.1177/0969141314548908 PMID: 25354870
53. Vieregge P, Stolze H, Klein C, Heberlein I. Gait quantitation in Parkinson’s disease? locomotor disability
and correlation to clinical rating scales. J Neural Transm 1997; 104: 237–48. https://doi.org/10.1007/
BF01273184 PMID: 9203085
54. Morris ME, Huxham FE, McGinley J, Iansek R. Gait disorders and gait rehabilitation in Parkinson’s dis-
ease. Adv Neurol 2001; 87: 347–61. PMID: 11347239
55. Morris M, Iansek R, McGinley J, Matyas T, Huxham F. Three-dimensional gait biomechanics in Parkin-
son’s disease: evidence for a centrally mediated amplitude regulation disorder. Mov Disord 2005; 20:
40–50. https://doi.org/10.1002/mds.20278 PMID: 15390033
56. Sofuwa O, Nieuwboer A, Desloovere K, Willems AM, Chavret F, Jonkers I. Quantitative Gait Analysis in
Parkinson’s Disease: Comparison With Healthy Control Group. Arch Phys Med Rehabil 2005; 86(5):
1007–1013. https://doi.org/10.1016/j.apmr.2004.08.012 PMID: 15895349
57. Serrao M, Pierelli F, Sinibaldi E, et al. Progressive Modular Rebalancing System and Visual Cueing for
Gait Rehabilitation in Parkinson’s Disease: A Pilot, Randomized, Controlled Trial With Crossover. Front
Neurol 2019; 10: 902. https://doi.org/10.3389/fneur.2019.00902 PMID: 31543859
58. Serrao M, Chini G, Caramanico G, et al. Prediction of Responsiveness of Gait Variables to Rehabilita-
tion Training in Parkinson’s Disease. Front Neurol 2019; 10: 826. https://doi.org/10.3389/fneur.2019.
00826 PMID: 31428039
59. Morris ME, Iansek R, Matyas TA, Summers JJ. The pathogenesis of gait hypokinesia in Parkinson’s dis-
ease. Brain 1994; 117(5): 1169–81. https://doi.org/10.1093/brain/117.5.1169 PMID: 7953597
60. Pawlitzki E, Schlenstedt C, Schmidt N, et al. Spatial orientation and postural control in patients with Par-
kinson’s disease. Gait Posture 2018; 60: 50–54. https://doi.org/10.1016/j.gaitpost.2017.11.011 PMID:
29153480
61. Durmus B, Baysal O, Altinayar S, Altay Z, Ersoy Y, Ozcan C. Lower extremity isokinetic muscle strength
in patients with Parkinson’s disease. J Clin Neurosci 2010; 17(7): 893–896. https://doi.org/10.1016/j.
jocn.2009.11.014 PMID: 20435478
62. Nocera JR, Horvat M, Ray CT. Impaired step up/over in persons with Parkinson’s disease. Adapt Phys
Activ Q 2010; 27(2): 87–95. https://doi.org/10.1123/apaq.27.2.87 PMID: 20440021
63. Djaldetti R, Shifrin A, Rogowski Z, Sprecher E, Melamed E, Yarnitsky D. Quantitative measurement of
pain sensation in patients with Parkinson disease. Neurology. 2004; 62(12): 2171–2175. https://doi.
org/10.1212/01.wnl.0000130455.38550.9d PMID: 15210877
64. Yoshii F, Moriya Y, Ohnuki T, Ryo M, Takahashi W. Postural deformities in Parkinson’s disease -Mutual
relationships among neck flexion, fore-bent, knee-bent and lateral-bent angles and correlations with
clinical predictors. J Clin Mov Disord 2016; 3: 1. https://doi.org/10.1186/s40734-016-0029-8 PMID:
26835153
65. Salarian A, Russmann H, Vingerhoets FJ, et al. Gait assessment in Parkinson’s disease: toward an
ambulatory system for long-term monitoring. IEEE Trans Biomed Eng. 2004; 51(8): 1434–1443. https://
doi.org/10.1109/TBME.2004.827933 PMID: 15311830
66. Perry J. Pathologic gait. Instr Course Lect 1990; 39: 325–331. PMID: 2186119
67. Begg RK, Sparrow WA. Ageing effects on knee and ankle joint angles at key events and phases of the
gait cycle. J Med Eng Technol 2006; 30(6): 382–389. https://doi.org/10.1080/03091900500445353
PMID: 17060166
68. Winter DA, Patla AE, Frank JS. Assessment of balance control in humans. Med Prog Technol 1990; 16
(1–2): 31–51. PMID: 2138696

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 15 / 16

PLOS ONE Neural Networks for Parkinson classification

69. Ashour R, Jankovic J. Joint and skeletal deformities in Parkinson’s disease, multiple system atrophy,
and progressive supranuclear palsy. Mov Disord 2006; 21(11): 1856–1863. https://doi.org/10.1002/
mds.21058 PMID: 16941460
70. Latt MD, Menz HB, Fung VS, Lord SR. Acceleration patterns of the head and pelvis during gait in older
people with Parkinson’s disease: a comparison of fallers and nonfallers. J Gerontol A Biol Sci Med Sci
2009; 64(6): 700–706. https://doi.org/10.1093/gerona/glp009 PMID: 19221191
71. Forsyth DR, Chia YC. How should Malaysia respond to its ageing society?. Med J Malaysia 2009; 64
(1): 46–50. PMID: 19852321
72. Barone P, Santangelo G, Amboni M, Pellecchia MT, Vitale C. Pisa syndrome in Parkinson’s disease
and parkinsonism: clinical features, pathophysiology, and treatment. Lancet Neurol 2016; 15(10):
1063–1074. https://doi.org/10.1016/S1474-4422(16)30173-9 PMID: 27571158
73. Mancini M, Horak FB, Zampieri C, Carlson-Kuhta P, Nutt JG, Chiari L. Trunk accelerometry reveals
postural instability in untreated Parkinson’s disease. Parkinsonism Relat Disord 2011; 17(7): 557–62.
https://doi.org/10.1016/j.parkreldis.2011.05.010 Epub 2011 Jun 8. PMID: 21641263
74. Florkowski CM. Sensitivity, specificity, receiver-operating characteristic (ROC) curves and likelihood
ratios: communicating the performance of diagnostic tests. Clin Biochem Rev. 2008; 29(1): S83–7.
PMID: 18852864
75. Jacobs JV, Horak FB. Abnormal proprioceptive-motor integration contributes to hypometric postural
responses of subjects with Parkinson’s disease. Neuroscience. 2006; 141: 999–1009. https://doi.org/
10.1016/j.neuroscience.2006.04.014 PMID: 16713110
76. Pozzo T, Berthoz A, Lefort L. Head stabilization during various locomotor tasks in humans. Exp Brain
Res 1990; 82: 97–106. https://doi.org/10.1007/BF00230842 PMID: 2257917
77. Prince F, Winter D, Stergiou P, Walt S. Anticipatory control of upper body balance during human loco-
motion. Gait Posture 1994; 2: 19–25. https://doi.org/10.1016/0966-6362(94)90013-2
78. Benatru I, Vaugoyeau M, Azulay JP. Postural disorders in Parkinson’s disease. Neurophysiol Clin Neu-
rophysiol 2008; 38:459–65. https://doi.org/10.1016/j.neucli.2008.07.006 PMID: 19026965
79. Bartolo M, Serrao M, Tassorelli C, et al. Fourweek trunk-specific rehabilitation treatment improves lat-
eral trunk flexion in Parkinson’s disease. Mov Disord 2010; 25: 325–31. https://doi.org/10.1002/mds.
23007 PMID: 20131386
80. Tassorelli C, De Icco R, Alfonsi E, et al. Botulinum toxin type A potentiates the effect of neuromotor
rehabilitation of Pisa syndrome in Parkinson disease: a placebo controlled study. Parkinsonism Relat
Disord 2014; 20(11): 1140–1144. https://doi.org/10.1016/j.parkreldis.2014.07.015 PMID: 25175601
81. Arendt-Nielsen L, Sinkjær T, Nielsen J, Kallesøe K. Electromyographic patterns and knee joint kinemat-
ics during walking at various speeds. J Electromyogr Kinesiol 1991; 1(2):89–95. https://doi.org/10.
1016/1050-6411(91)90002-M PMID: 20870498
82. Hebenstreit F, Leibold A, Krinner S, Welsch G, Lochmann M, Eskofier BM. Effect of walking speed on
gait sub phase durations. Hum Mov Sci. 2015; 43: 118–124. https://doi.org/10.1016/j.humov.2015.07.
009 PMID: 26256534
83. Romkes J, Bracht-Schweizer K. The effects of walking speed on upper body kinematics during gait in
healthy subjects. Gait Posture 2017; 54: 304–310. https://doi.org/10.1016/j.gaitpost.2017.03.025
PMID: 28395170
84. Schreiber C, Armand S, Moissenet F. Influence of normative data’s walking speed on the computation
of conventional gait indices. J Biomech 2018; 76: 68–73. https://doi.org/10.1016/j.jbiomech.2018.05.
022 PMID: 29853318
85. Fukuchi CA, Fukuchi RK, Duarte M. Effects of walking speed on gait biomechanics in healthy partici-
pants: a systematic review and meta-analysis. Syst Rev 2019; 8(1): 153. https://doi.org/10.1186/
s13643-019-1063-z PMID: 31248456
86. Louie S, Koop MM, Frenklach A, Bronte-Stewart. Quantitative lateralized measures of bradykinesia at
different stages of Parkinson’s disease: the role of the less affected side. H Mov Disord 2009; 24(13):
1991–7. https://doi.org/10.1002/mds.22741 PMID: 19672996
87. Roiz Rde M, Cacho EW, Pazinatto MM, Reis JG, Cliquet A Jr, Barasnevicius-Quagliato EM. Gait analy-
sis comparing Parkinson’s disease with healthy elderly subjects. Arq Neuropsiquiatr. 2010; 68(1): 81–
6. https://doi.org/10.1590/s0004-282x2010000100018 PMID: 20339659
88. Rehman RZU, Del Din S, Guan Y, Yarnall AJ, Shi JQ, Rochester L. Selecting Clinically Relevant Gait
Characteristics for Classification of Early Parkinson’s Disease: A Comprehensive Machine Learning
Approach. Sci Rep. 2019; 9(1): 17269. https://doi.org/10.1038/s41598-019-53656-7 PMID: 31754175
89. Gao C, Sun H, Wang T, Tang M, Bohnen NI, Müller MLTM, et al. Model-based and Model-free Machine
Learning Techniques for Diagnostic Prediction and Classification of Clinical Outcomes in Parkinson’s
Disease. Sci Rep. 2018; 8(1): 7129. https://doi.org/10.1038/s41598-018-24783-4 PMID: 29740058

PLOS ONE | https://doi.org/10.1371/journal.pone.0244396 February 19, 2021 16 / 16